Image segmentation techniques are widely used for segmenting medical images and determining contours between anatomical structures within the images. For example, in radiation therapy, automatic segmentation of organs is usually performed to reduce contouring time, and improve contour accuracy and consistency over various hospitals. However, automated segmentation remains to be a very difficult task on medical images having lower image quality, such as some computer tomography (CT) or cone-beam computer tomography (CBCT) images that may be used to treat cancer patients. For example, such CT or CBCT images are known to have lower contrast and little textures for most soft tissue structures. Therefore, conventional image segmentation methods based primarily on image contrast often fail to find an accurate contour between the background and anatomical structures (e.g., organs or tumors) of interest, or between different anatomical structures in a medical image.
FIG. 1 illustrates an exemplary three-dimensional (3D) CT image from a typical prostate cancer patient. Illustration (A) shows a pelvic region of the patient in a 3D view, which includes the patient's bladder, prostate, and rectum. Images (B), (C), and (D) are axial, sagittal and coronal views from a 3D CT image of this pelvic region. As shown in images (B), (C), and (D), most part of the patient's prostate boundary is not visible. That is, one cannot readily distinguish the prostate from other anatomical structures or determine a contour for the prostate. In comparison, images (E), (F), and (G) show the expected prostate contour on the same 3D CT image. Therefore, conventional image segmentation methods based on solely the contrast and textures presented in the image will likely fail when used to segment this exemplary 3D CT image.
Recent developments in machine learning techniques make improved image segmentation on lower quality images possible. For example, supervised learning algorithms can “train” the machines or computers to predict which anatomical structure each pixel or voxel of a medical Image should belong to. Such prediction usually uses features of the pixel or voxel as inputs. Therefore, the performance of the segmentation highly depends on the type of features available. To date, most learning-based image segmentation methods are based primarily on local image features such as image intensities, image textures, etc. As a result, these segmentation methods are still suboptimal for lower quality images, such as the 3D CT image shown in FIG. 1.
Accordingly, there is a need to design more appropriate features for learning-based auto-segmentation methods in order to improve segmentation performance on medical images in radiation therapy or related fields.